Optimal potato (Solanum tuberosum L.) yield depends heavily on precise assessment of irrigation water quality, yet conventional laboratory analyses often provide only generic potability indices that lack crop-specific, model-driven decision support. To bridge this gap, this study introduces a rapid and accurate classification framework that leverages routine physicochemical parameters (pH, hardness, solids, chloramines, sulphate, organic carbon, trihalomethanes, and turbidity). We adapt Facebook’s Regular Network (RegNet) architecture to structured water-quality data and employ the Ninja Optimization Algorithm (NinjaOA) for systematic hyperparameter tuning. Experiments were conducted using the publicly available Kaggle “Potato Crop Water Quality Parameters” dataset, with a binary “Check” flag indicating suitability. Baseline comparisons against leading tabular models—including Deep Gradient Boosting Machine (DeepGBM), Self-Attention and Intersample Transformer (SAINT), Neural Oblivious Decision Ensembles (NODE), Tabular Network (TabNet), and Deep Learning Network (DLN)—showed that untuned RegNet already delivered the best performance (accuracy 0.8679, $${F}_{{1}}$$-score 0.8662). After optimization, NinjaOA+RegNet achieved accuracy 0.9798 and $${F}_{{1}}$$-score 0.9796, surpassing the next-best optimizer, genetic algorithm (GA), by 1.18 points in accuracy and 1.20 in $${F}_{{1}}$$, while substantially outperforming the untuned baseline. These results establish NinjaOA-tuned RegNet as a high-performance benchmark for potato irrigation water suitability classification, underscoring the synergy between deep backbone adaptation and metaheuristic hyperparameter optimization. Its modular design supports transferability to other crops and water sources, while reliance on standard water quality parameters enhances feasibility for real-world precision agriculture deployments.
Full publication URL
